service: add count per lcore
[dpdk.git] / lib / librte_eal / common / rte_service.c
1 /*-
2  *   BSD LICENSE
3  *
4  *   Copyright(c) 2017 Intel Corporation. All rights reserved.
5  *   All rights reserved.
6  *
7  *   Redistribution and use in source and binary forms, with or without
8  *   modification, are permitted provided that the following conditions
9  *   are met:
10  *
11  *     * Redistributions of source code must retain the above copyright
12  *       notice, this list of conditions and the following disclaimer.
13  *     * Redistributions in binary form must reproduce the above copyright
14  *       notice, this list of conditions and the following disclaimer in
15  *       the documentation and/or other materials provided with the
16  *       distribution.
17  *     * Neither the name of Intel Corporation nor the names of its
18  *       contributors may be used to endorse or promote products derived
19  *       from this software without specific prior written permission.
20  *
21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
32  */
33
34 #include <stdio.h>
35 #include <unistd.h>
36 #include <inttypes.h>
37 #include <limits.h>
38 #include <string.h>
39 #include <dirent.h>
40
41 #include <rte_service.h>
42 #include "include/rte_service_component.h"
43
44 #include <rte_eal.h>
45 #include <rte_lcore.h>
46 #include <rte_common.h>
47 #include <rte_debug.h>
48 #include <rte_cycles.h>
49 #include <rte_atomic.h>
50 #include <rte_memory.h>
51 #include <rte_malloc.h>
52
53 #define RTE_SERVICE_NUM_MAX 64
54
55 #define SERVICE_F_REGISTERED    (1 << 0)
56 #define SERVICE_F_STATS_ENABLED (1 << 1)
57
58 /* runstates for services and lcores, denoting if they are active or not */
59 #define RUNSTATE_STOPPED 0
60 #define RUNSTATE_RUNNING 1
61
62 /* internal representation of a service */
63 struct rte_service_spec_impl {
64         /* public part of the struct */
65         struct rte_service_spec spec;
66
67         /* atomic lock that when set indicates a service core is currently
68          * running this service callback. When not set, a core may take the
69          * lock and then run the service callback.
70          */
71         rte_atomic32_t execute_lock;
72
73         /* API set/get-able variables */
74         int32_t runstate;
75         uint8_t internal_flags;
76
77         /* per service statistics */
78         uint32_t num_mapped_cores;
79         uint64_t calls;
80         uint64_t cycles_spent;
81 } __rte_cache_aligned;
82
83 /* the internal values of a service core */
84 struct core_state {
85         /* map of services IDs are run on this core */
86         uint64_t service_mask;
87         uint8_t runstate; /* running or stopped */
88         uint8_t is_service_core; /* set if core is currently a service core */
89
90         /* extreme statistics */
91         uint64_t calls_per_service[RTE_SERVICE_NUM_MAX];
92 } __rte_cache_aligned;
93
94 static uint32_t rte_service_count;
95 static struct rte_service_spec_impl *rte_services;
96 static struct core_state *lcore_states;
97 static uint32_t rte_service_library_initialized;
98
99 int32_t rte_service_init(void)
100 {
101         if (rte_service_library_initialized) {
102                 printf("service library init() called, init flag %d\n",
103                         rte_service_library_initialized);
104                 return -EALREADY;
105         }
106
107         rte_services = rte_calloc("rte_services", RTE_SERVICE_NUM_MAX,
108                         sizeof(struct rte_service_spec_impl),
109                         RTE_CACHE_LINE_SIZE);
110         if (!rte_services) {
111                 printf("error allocating rte services array\n");
112                 return -ENOMEM;
113         }
114
115         lcore_states = rte_calloc("rte_service_core_states", RTE_MAX_LCORE,
116                         sizeof(struct core_state), RTE_CACHE_LINE_SIZE);
117         if (!lcore_states) {
118                 printf("error allocating core states array\n");
119                 return -ENOMEM;
120         }
121
122         int i;
123         int count = 0;
124         struct rte_config *cfg = rte_eal_get_configuration();
125         for (i = 0; i < RTE_MAX_LCORE; i++) {
126                 if (lcore_config[i].core_role == ROLE_SERVICE) {
127                         if ((unsigned int)i == cfg->master_lcore)
128                                 continue;
129                         rte_service_lcore_add(i);
130                         count++;
131                 }
132         }
133
134         rte_service_library_initialized = 1;
135         return 0;
136 }
137
138 /* returns 1 if service is registered and has not been unregistered
139  * Returns 0 if service never registered, or has been unregistered
140  */
141 static inline int
142 service_valid(uint32_t id)
143 {
144         return !!(rte_services[id].internal_flags & SERVICE_F_REGISTERED);
145 }
146
147 /* returns 1 if statistics should be colleced for service
148  * Returns 0 if statistics should not be collected for service
149  */
150 static inline int
151 service_stats_enabled(struct rte_service_spec_impl *impl)
152 {
153         return !!(impl->internal_flags & SERVICE_F_STATS_ENABLED);
154 }
155
156 static inline int
157 service_mt_safe(struct rte_service_spec_impl *s)
158 {
159         return s->spec.capabilities & RTE_SERVICE_CAP_MT_SAFE;
160 }
161
162 int32_t rte_service_set_stats_enable(struct rte_service_spec *service,
163                                   int32_t enabled)
164 {
165         struct rte_service_spec_impl *impl =
166                 (struct rte_service_spec_impl *)service;
167         if (!impl)
168                 return -EINVAL;
169
170         if (enabled)
171                 impl->internal_flags |= SERVICE_F_STATS_ENABLED;
172         else
173                 impl->internal_flags &= ~(SERVICE_F_STATS_ENABLED);
174
175         return 0;
176 }
177
178 uint32_t
179 rte_service_get_count(void)
180 {
181         return rte_service_count;
182 }
183
184 struct rte_service_spec *
185 rte_service_get_by_id(uint32_t id)
186 {
187         struct rte_service_spec *service = NULL;
188         if (id < rte_service_count)
189                 service = (struct rte_service_spec *)&rte_services[id];
190
191         return service;
192 }
193
194 struct rte_service_spec *rte_service_get_by_name(const char *name)
195 {
196         struct rte_service_spec *service = NULL;
197         int i;
198         for (i = 0; i < RTE_SERVICE_NUM_MAX; i++) {
199                 if (service_valid(i) &&
200                                 strcmp(name, rte_services[i].spec.name) == 0) {
201                         service = (struct rte_service_spec *)&rte_services[i];
202                         break;
203                 }
204         }
205
206         return service;
207 }
208
209 const char *
210 rte_service_get_name(const struct rte_service_spec *service)
211 {
212         return service->name;
213 }
214
215 int32_t
216 rte_service_probe_capability(const struct rte_service_spec *service,
217                              uint32_t capability)
218 {
219         return service->capabilities & capability;
220 }
221
222 int32_t
223 rte_service_is_running(const struct rte_service_spec *spec)
224 {
225         const struct rte_service_spec_impl *impl =
226                 (const struct rte_service_spec_impl *)spec;
227         if (!impl)
228                 return -EINVAL;
229
230         return (impl->runstate == RUNSTATE_RUNNING) &&
231                 (impl->num_mapped_cores > 0);
232 }
233
234 int32_t
235 rte_service_register(const struct rte_service_spec *spec)
236 {
237         uint32_t i;
238         int32_t free_slot = -1;
239
240         if (spec->callback == NULL || strlen(spec->name) == 0)
241                 return -EINVAL;
242
243         for (i = 0; i < RTE_SERVICE_NUM_MAX; i++) {
244                 if (!service_valid(i)) {
245                         free_slot = i;
246                         break;
247                 }
248         }
249
250         if ((free_slot < 0) || (i == RTE_SERVICE_NUM_MAX))
251                 return -ENOSPC;
252
253         struct rte_service_spec_impl *s = &rte_services[free_slot];
254         s->spec = *spec;
255         s->internal_flags |= SERVICE_F_REGISTERED;
256
257         rte_smp_wmb();
258         rte_service_count++;
259
260         return 0;
261 }
262
263 int32_t
264 rte_service_unregister(struct rte_service_spec *spec)
265 {
266         struct rte_service_spec_impl *s = NULL;
267         struct rte_service_spec_impl *spec_impl =
268                 (struct rte_service_spec_impl *)spec;
269
270         uint32_t i;
271         uint32_t service_id;
272         for (i = 0; i < RTE_SERVICE_NUM_MAX; i++) {
273                 if (&rte_services[i] == spec_impl) {
274                         s = spec_impl;
275                         service_id = i;
276                         break;
277                 }
278         }
279
280         if (!s)
281                 return -EINVAL;
282
283         rte_service_count--;
284         rte_smp_wmb();
285
286         s->internal_flags &= ~(SERVICE_F_REGISTERED);
287
288         for (i = 0; i < RTE_MAX_LCORE; i++)
289                 lcore_states[i].service_mask &= ~(UINT64_C(1) << service_id);
290
291         memset(&rte_services[service_id], 0,
292                         sizeof(struct rte_service_spec_impl));
293
294         return 0;
295 }
296
297 int32_t
298 rte_service_start(struct rte_service_spec *service)
299 {
300         struct rte_service_spec_impl *s =
301                 (struct rte_service_spec_impl *)service;
302         s->runstate = RUNSTATE_RUNNING;
303         rte_smp_wmb();
304         return 0;
305 }
306
307 int32_t
308 rte_service_stop(struct rte_service_spec *service)
309 {
310         struct rte_service_spec_impl *s =
311                 (struct rte_service_spec_impl *)service;
312         s->runstate = RUNSTATE_STOPPED;
313         rte_smp_wmb();
314         return 0;
315 }
316
317 static int32_t
318 rte_service_runner_func(void *arg)
319 {
320         RTE_SET_USED(arg);
321         uint32_t i;
322         const int lcore = rte_lcore_id();
323         struct core_state *cs = &lcore_states[lcore];
324
325         while (lcore_states[lcore].runstate == RUNSTATE_RUNNING) {
326                 const uint64_t service_mask = cs->service_mask;
327                 for (i = 0; i < rte_service_count; i++) {
328                         struct rte_service_spec_impl *s = &rte_services[i];
329                         if (s->runstate != RUNSTATE_RUNNING ||
330                                         !(service_mask & (UINT64_C(1) << i)))
331                                 continue;
332
333                         /* check do we need cmpset, if MT safe or <= 1 core
334                          * mapped, atomic ops are not required.
335                          */
336                         const int need_cmpset = !((service_mt_safe(s) == 0) &&
337                                                 (s->num_mapped_cores > 1));
338                         uint32_t *lock = (uint32_t *)&s->execute_lock;
339
340                         if (need_cmpset || rte_atomic32_cmpset(lock, 0, 1)) {
341                                 void *userdata = s->spec.callback_userdata;
342
343                                 if (service_stats_enabled(s)) {
344                                         uint64_t start = rte_rdtsc();
345                                         s->spec.callback(userdata);
346                                         uint64_t end = rte_rdtsc();
347                                         s->cycles_spent += end - start;
348                                         cs->calls_per_service[i]++;
349                                         s->calls++;
350                                 } else
351                                         s->spec.callback(userdata);
352
353                                 if (need_cmpset)
354                                         rte_atomic32_clear(&s->execute_lock);
355                         }
356                 }
357
358                 rte_smp_rmb();
359         }
360
361         lcore_config[lcore].state = WAIT;
362
363         return 0;
364 }
365
366 int32_t
367 rte_service_lcore_count(void)
368 {
369         int32_t count = 0;
370         uint32_t i;
371         for (i = 0; i < RTE_MAX_LCORE; i++)
372                 count += lcore_states[i].is_service_core;
373         return count;
374 }
375
376 int32_t
377 rte_service_lcore_list(uint32_t array[], uint32_t n)
378 {
379         uint32_t count = rte_service_lcore_count();
380         if (count > n)
381                 return -ENOMEM;
382
383         if (!array)
384                 return -EINVAL;
385
386         uint32_t i;
387         uint32_t idx = 0;
388         for (i = 0; i < RTE_MAX_LCORE; i++) {
389                 struct core_state *cs = &lcore_states[i];
390                 if (cs->is_service_core) {
391                         array[idx] = i;
392                         idx++;
393                 }
394         }
395
396         return count;
397 }
398
399 int32_t
400 rte_service_lcore_count_services(uint32_t lcore)
401 {
402         if (lcore >= RTE_MAX_LCORE)
403                 return -EINVAL;
404
405         struct core_state *cs = &lcore_states[lcore];
406         if (!cs->is_service_core)
407                 return -ENOTSUP;
408
409         return __builtin_popcountll(cs->service_mask);
410 }
411
412 int32_t
413 rte_service_start_with_defaults(void)
414 {
415         /* create a default mapping from cores to services, then start the
416          * services to make them transparent to unaware applications.
417          */
418         uint32_t i;
419         int ret;
420         uint32_t count = rte_service_get_count();
421
422         int32_t lcore_iter = 0;
423         uint32_t ids[RTE_MAX_LCORE];
424         int32_t lcore_count = rte_service_lcore_list(ids, RTE_MAX_LCORE);
425
426         if (lcore_count == 0)
427                 return -ENOTSUP;
428
429         for (i = 0; (int)i < lcore_count; i++)
430                 rte_service_lcore_start(ids[i]);
431
432         for (i = 0; i < count; i++) {
433                 struct rte_service_spec *s = rte_service_get_by_id(i);
434                 if (!s)
435                         return -EINVAL;
436
437                 /* do 1:1 core mapping here, with each service getting
438                  * assigned a single core by default. Adding multiple services
439                  * should multiplex to a single core, or 1:1 if there are the
440                  * same amount of services as service-cores
441                  */
442                 ret = rte_service_enable_on_lcore(s, ids[lcore_iter]);
443                 if (ret)
444                         return -ENODEV;
445
446                 lcore_iter++;
447                 if (lcore_iter >= lcore_count)
448                         lcore_iter = 0;
449
450                 ret = rte_service_start(s);
451                 if (ret)
452                         return -ENOEXEC;
453         }
454
455         return 0;
456 }
457
458 static int32_t
459 service_update(struct rte_service_spec *service, uint32_t lcore,
460                 uint32_t *set, uint32_t *enabled)
461 {
462         uint32_t i;
463         int32_t sid = -1;
464
465         for (i = 0; i < RTE_SERVICE_NUM_MAX; i++) {
466                 if ((struct rte_service_spec *)&rte_services[i] == service &&
467                                 service_valid(i)) {
468                         sid = i;
469                         break;
470                 }
471         }
472
473         if (sid == -1 || lcore >= RTE_MAX_LCORE)
474                 return -EINVAL;
475
476         if (!lcore_states[lcore].is_service_core)
477                 return -EINVAL;
478
479         uint64_t sid_mask = UINT64_C(1) << sid;
480         if (set) {
481                 if (*set) {
482                         lcore_states[lcore].service_mask |= sid_mask;
483                         rte_services[sid].num_mapped_cores++;
484                 } else {
485                         lcore_states[lcore].service_mask &= ~(sid_mask);
486                         rte_services[sid].num_mapped_cores--;
487                 }
488         }
489
490         if (enabled)
491                 *enabled = (lcore_states[lcore].service_mask & (sid_mask));
492
493         rte_smp_wmb();
494
495         return 0;
496 }
497
498 int32_t rte_service_get_enabled_on_lcore(struct rte_service_spec *service,
499                                         uint32_t lcore)
500 {
501         uint32_t enabled;
502         int ret = service_update(service, lcore, 0, &enabled);
503         if (ret == 0)
504                 return enabled;
505         return -EINVAL;
506 }
507
508 int32_t
509 rte_service_enable_on_lcore(struct rte_service_spec *service, uint32_t lcore)
510 {
511         uint32_t on = 1;
512         return service_update(service, lcore, &on, 0);
513 }
514
515 int32_t
516 rte_service_disable_on_lcore(struct rte_service_spec *service, uint32_t lcore)
517 {
518         uint32_t off = 0;
519         return service_update(service, lcore, &off, 0);
520 }
521
522 int32_t rte_service_lcore_reset_all(void)
523 {
524         /* loop over cores, reset all to mask 0 */
525         uint32_t i;
526         for (i = 0; i < RTE_MAX_LCORE; i++) {
527                 lcore_states[i].service_mask = 0;
528                 lcore_states[i].is_service_core = 0;
529                 lcore_states[i].runstate = RUNSTATE_STOPPED;
530         }
531         for (i = 0; i < RTE_SERVICE_NUM_MAX; i++)
532                 rte_services[i].num_mapped_cores = 0;
533
534         rte_smp_wmb();
535
536         return 0;
537 }
538
539 static void
540 set_lcore_state(uint32_t lcore, int32_t state)
541 {
542         /* mark core state in hugepage backed config */
543         struct rte_config *cfg = rte_eal_get_configuration();
544         cfg->lcore_role[lcore] = state;
545
546         /* mark state in process local lcore_config */
547         lcore_config[lcore].core_role = state;
548
549         /* update per-lcore optimized state tracking */
550         lcore_states[lcore].is_service_core = (state == ROLE_SERVICE);
551 }
552
553 int32_t
554 rte_service_lcore_add(uint32_t lcore)
555 {
556         if (lcore >= RTE_MAX_LCORE)
557                 return -EINVAL;
558         if (lcore_states[lcore].is_service_core)
559                 return -EALREADY;
560
561         set_lcore_state(lcore, ROLE_SERVICE);
562
563         /* ensure that after adding a core the mask and state are defaults */
564         lcore_states[lcore].service_mask = 0;
565         lcore_states[lcore].runstate = RUNSTATE_STOPPED;
566
567         rte_smp_wmb();
568         return 0;
569 }
570
571 int32_t
572 rte_service_lcore_del(uint32_t lcore)
573 {
574         if (lcore >= RTE_MAX_LCORE)
575                 return -EINVAL;
576
577         struct core_state *cs = &lcore_states[lcore];
578         if (!cs->is_service_core)
579                 return -EINVAL;
580
581         if (cs->runstate != RUNSTATE_STOPPED)
582                 return -EBUSY;
583
584         set_lcore_state(lcore, ROLE_RTE);
585
586         rte_smp_wmb();
587         return 0;
588 }
589
590 int32_t
591 rte_service_lcore_start(uint32_t lcore)
592 {
593         if (lcore >= RTE_MAX_LCORE)
594                 return -EINVAL;
595
596         struct core_state *cs = &lcore_states[lcore];
597         if (!cs->is_service_core)
598                 return -EINVAL;
599
600         if (cs->runstate == RUNSTATE_RUNNING)
601                 return -EALREADY;
602
603         /* set core to run state first, and then launch otherwise it will
604          * return immediately as runstate keeps it in the service poll loop
605          */
606         lcore_states[lcore].runstate = RUNSTATE_RUNNING;
607
608         int ret = rte_eal_remote_launch(rte_service_runner_func, 0, lcore);
609         /* returns -EBUSY if the core is already launched, 0 on success */
610         return ret;
611 }
612
613 int32_t
614 rte_service_lcore_stop(uint32_t lcore)
615 {
616         if (lcore >= RTE_MAX_LCORE)
617                 return -EINVAL;
618
619         if (lcore_states[lcore].runstate == RUNSTATE_STOPPED)
620                 return -EALREADY;
621
622         uint32_t i;
623         for (i = 0; i < RTE_SERVICE_NUM_MAX; i++) {
624                 int32_t enabled =
625                         lcore_states[i].service_mask & (UINT64_C(1) << i);
626                 int32_t service_running = rte_services[i].runstate !=
627                                                 RUNSTATE_STOPPED;
628                 int32_t only_core = rte_services[i].num_mapped_cores == 1;
629
630                 /* if the core is mapped, and the service is running, and this
631                  * is the only core that is mapped, the service would cease to
632                  * run if this core stopped, so fail instead.
633                  */
634                 if (enabled && service_running && only_core)
635                         return -EBUSY;
636         }
637
638         lcore_states[lcore].runstate = RUNSTATE_STOPPED;
639
640         return 0;
641 }
642
643 static void
644 rte_service_dump_one(FILE *f, struct rte_service_spec_impl *s,
645                      uint64_t all_cycles, uint32_t reset)
646 {
647         /* avoid divide by zero */
648         if (all_cycles == 0)
649                 all_cycles = 1;
650
651         int calls = 1;
652         if (s->calls != 0)
653                 calls = s->calls;
654
655         fprintf(f, "  %s: stats %d\tcalls %"PRIu64"\tcycles %"
656                         PRIu64"\tavg: %"PRIu64"\n",
657                         s->spec.name, service_stats_enabled(s), s->calls,
658                         s->cycles_spent, s->cycles_spent / calls);
659
660         if (reset) {
661                 s->cycles_spent = 0;
662                 s->calls = 0;
663         }
664 }
665
666 static void
667 service_dump_calls_per_lcore(FILE *f, uint32_t lcore, uint32_t reset)
668 {
669         uint32_t i;
670         struct core_state *cs = &lcore_states[lcore];
671
672         fprintf(f, "%02d\t", lcore);
673         for (i = 0; i < RTE_SERVICE_NUM_MAX; i++) {
674                 if (!service_valid(i))
675                         continue;
676                 fprintf(f, "%"PRIu64"\t", cs->calls_per_service[i]);
677                 if (reset)
678                         cs->calls_per_service[i] = 0;
679         }
680         fprintf(f, "\n");
681 }
682
683 int32_t rte_service_dump(FILE *f, struct rte_service_spec *service)
684 {
685         uint32_t i;
686
687         uint64_t total_cycles = 0;
688         for (i = 0; i < rte_service_count; i++) {
689                 if (!service_valid(i))
690                         continue;
691                 total_cycles += rte_services[i].cycles_spent;
692         }
693
694         if (service) {
695                 struct rte_service_spec_impl *s =
696                         (struct rte_service_spec_impl *)service;
697                 fprintf(f, "Service %s Summary\n", s->spec.name);
698                 uint32_t reset = 0;
699                 rte_service_dump_one(f, s, total_cycles, reset);
700                 return 0;
701         }
702
703         fprintf(f, "Services Summary\n");
704         for (i = 0; i < rte_service_count; i++) {
705                 uint32_t reset = 1;
706                 rte_service_dump_one(f, &rte_services[i], total_cycles, reset);
707         }
708
709         fprintf(f, "Service Cores Summary\n");
710         for (i = 0; i < RTE_MAX_LCORE; i++) {
711                 if (lcore_config[i].core_role != ROLE_SERVICE)
712                         continue;
713
714                 uint32_t reset = 0;
715                 service_dump_calls_per_lcore(f, i, reset);
716         }
717
718         return 0;
719 }